Methods of Measuring Usual

Dietary Intake for Risk

Assessment

Amy F Subar, PhD, MPH, RDApplied Research Program

Risk Factor Monitoring and Methods Branch

Objectives

• Describe major dietary assessment

approaches used in population studies

• Highlight strengths and weaknesses

• Describe and discuss innovative methods– Collecting data

– Analyzing data

I gather inaccurate data for a living. Luckily no one realizes it.

Your glass is half full.

Main Dietary Assessment Methods

in Population Studies

• Food records or diaries

• 24-Hour dietary recalls

• Food frequency questionnaires (FFQs)

• Brief methods

Food Diaries or Records

• Variations in approach

– Trained or untrained respondents

– Detailed review or not

– Highly standardized coding rules or not

– Layout of record, instructions

– Development of electronic methods: PDA, cell

phones

– Quality and completeness of nutrient database

Food Diaries or Records:

Strengths

• Intake is quantified, detailed

• Could be relatively accurate

• Data are rich

– Nutrients

– Cooking practices

– Meal and eating frequency

Food Records/Diaries:

Weaknesses• Recording influences diet (reactive tool)

– Biased measurement

• Requires literacy

• High respondent and investigator burden

• Multiple days required to estimate usual intake

• Sample selection bias

• Completion worse over time

• Underreporting is typical– Worse with overweight/obese

24-Hour Dietary Recalls

• Variations in approach

– Training of interviewers

– Standardization of probing questions

– Multiple passes through the day

– Computer vs. paper/pencil administration

– In-person vs. telephone

– Portion size models or measurement aids

– Quality and completeness of nutrient database

– Development of self-administered automated

systems

24-Hour Dietary Recall: Strengths

• Intake is quantified, detailed

• Data are rich– Nutrients

– Cooking practices

– Meal and eating frequency

• Does not affect eating behavior

• Lower sample selection bias than record– Literacy not required (if interviewer administered)

– Respondent burden relatively low

24-Hour Dietary Recall: Weaknesses

• Imperfect knowledge and memory

• Multiple days required to estimate usual intake

• Costly to administer

– Highly trained interviewers

• Some evidence that reporting declines with

multiple administrations

• Underreporting is typical

– Worse with overweight/obese

Food Frequency Questionnaire (FFQ)

• Variations in approach

– Number of foods, clarity of questions

– Portion size questions: pictures vs. text description

– Time frame

– Development of food list, nutrient database

– Type of administration: computer vs. paper

– Specificity to population of interest

– Supplement intake

– Food preparation

FFQ: Strengths

• Low respondent burden

• Attempts to estimate “usual” individual intake

of foods with one administration

• Information on total diet queried

– Depends on food list

• Does not affect eating behavior

• Low cost of administering/processing

– Machine-readable scanned forms

– Computer-generated data

FFQ: Weaknesses

• Lacks detail

– Finite food list

– Details of cooking methods, specific food types lost

• Cognitively complex

• Requires literacy

• Severe measurement error: prone to bias

• Different FFQs can behave quite differently

• Different populations respond differently

Biomarkers

• Recovery: represent 1:1 intake exposure – Very few available

• Concentration: reflect the biological response to intake– Cannot be used to assess amount consumed

– Correlate to amount consumed

– May reflect short or long term intakes

– Some are appropriate for risk assessment

• Tightly controlled: no relationship to intake

Isn’t There ANYTHING

Better Than This?

Innovative Ideas in Dietary

Assessment

• Web-based, automated FFQs

• Self-administered, automated, web-based, 24-

hour dietary recalls

• Real time data collection of food records

• Voice recognition software

• Combining methods

– FFQ with 24-hour recalls

Web-Based, Automated FFQs

• Advantages beyond paper and pencil:– Cleaner data

– May take less time to complete

• Disadvantages beyond paper and pencil– (It’s still an FFQ)

– Computer access and related technology issues

NCI’s Vision for an Automated Self-

Administered 24-Hour Recall (ASA24)

• 24-HRs that are automated AND self-administered:– Complete system for probing, coding, analysis

– Accessible on the Web, publicly available

– Easily updated

• Adaptable to other languages

• Modeled after dietary surveillance systems in NHANES (AMPM)

• Multiple 24-HRs could be collected for minimal cost

ASA24

Next Steps

• Alpha (α) prototypes

• Conduct cognitive testing

• Expected completion: Summer 2007

• Validation planned

NCI Method to Estimate

Usual Intake

• New statistical method to estimate

– Usual nutrient and food intake distributions for

dietary surveillance or risk assessment

– Individual usual nutrient and food intake for

assessing the relationship between food intake

and disease risk

Definition of Usual Intake

• Theoretical long-run average daily

intake of a dietary component

• Common operational definition: average

daily intake over past year

Usual Intake Distribution

for the Population• Needed to estimate proportion of the

population above or below a particular

cutoff

1-day vs. Usual Intake (UI) Distributions

Usual Intakes

One-day Intakes

Evolution in Estimating Usual

Nutrient Intake Distributions

• Generally uses 24-HR and/or records

• Average of a few single-day measurements

• Statistical modeling

– National Research Council method

– Iowa State University method

Statistical Modeling

• Removes day-to-day variability from 24-HRs

• Approach well-established for:– Most nutrients

– Foods consumed daily by nearly everyone (such as total fruits and vegetables)

• Approach less well-established for episodically consumed foods

Estimating UI Distributions of

Episodically Consumed Foods

• Small number of 24-HRs makes this difficult

– large number of “zero” intakes observed

Challenge: Zero Single-Day Intakes

Spike

at

Zero

Skewed

Distribution

Men

Source: EATS

Response: Decompose Usual Intake

Probability × Amount=Usual Intake

• Relative frequency of nonzero 24-HRs provides information about the distribution of consumption probability

• Amounts on consumption days always positive, so apply “nutrient” methods for distribution of usual consumption-day intake

• Combine both distributions to get distribution of usual intake

The NCI Method For Foods

InterceptI + SlopeI × CovariateI

+ Person-specific EffectI

=logit(24-HR intake)

InterceptII + SlopeII × CovariateII

+ Person-specific EffectII + ErrorII

=Transformed(24-HR intake)

Part I (Probability): Logistic regression with person-specific effects

Part II (Amount): Nonlinear mixed model as for nutrients

Both parts estimated simultaneously (advantage over ISU method)

• Random effects can be correlated

• Two parts of the model can share covariates

Two-Part Nonlinear Mixed Model

NCI Method: Efficiently Accounts for

Effects of Covariates on Usual Intake

• Separate estimation of usual intake distributions for subgroups may be impractical/inefficient

• Use of covariates in NCI method allows direct evaluation of covariate effects on usual intake, in addition to correction for measurement error

Food Propensity Questionnaire

(FPQ)

• FFQ that does not query portion size

• Directly queries long-term frequency of

intake

• Relatively low respondent/interviewer

burden

• Can be used as a covariate in NCI model

Rationale for Using an FPQ as a

Covariate

• Responses to FPQ are related to information from 24-HR– Probability

– Amount

• Based on these relationships, FPQ responses act as covariates in model estimating individual usual intake from 24-HRs

Can We Trust the FPQ?

• Concerns about bias in FFQs when used to measure absolute intake

• In this case, FPQ being used:– In conjunction with another instrument

– As covariate in model to predict probability and amount of food consumed

– Nutrient or food group estimates from the FPQ are not used

Relationships Between

Diet and Outcomes

• Challenge: usual intake is unobservable

– have 24-HR intakes or FFQs instead

• Typical Response: Fit disease model to

average of 24-HR intakes or FFQ

• NCI method suggests that at least 2 recalls

and a FPQ might be useful

Effect of Measurement Error

-3 -2 -1 0 1 2 3

-3-2

-10

12

3

Observed (red) vs true (blue) dataSource: Simulated data

Effect of Measurement Error

• Typically measurement error causes two things:

– bias in the estimated exposure effect (flattened or

attenuated true slope)

– more variation about the flattened line, thus loss of

statistical power for testing significance of the

exposure effect

Summary: NCI Method

• Important assumption that 24-HR is an unbiased instrument for measuring usual intake

• Misreporting of energy on 24-HR suggests misreporting of some foods

• For foods reported with bias on 24-HR, estimates of usual intake will be biased

• FPQ may improve estimation

• Development of user-friendly software for this purpose

• Requires collaboration between statisticians and nutritionists

• October Journal of the American Dietetic Assn: 3 papers on this work: Dodd et al, Subar et al, Tooze et al

Conclusion

• A good estimate of usual intake is necessary for

risk assessment

• Usual intake cannot be observed

– Can be modeled

GOOD LUCK TO US ALL!